Check out "General Adversarial Networks". It's basically training a two-part AI via an arms race; one part is trying to fake out the other, the other is trying to detect the fake.
Another way to describe GANs is that the loss function, which used to be hand-made, now is a whole separate neural net and can be learned bottom-up instead of dictated top-down.
There was a time when features were hand-made. Now we have deep learning and finding good features is automatic. GANs automate another part of the process.